Author: Robert Hyatt
Date: 08:02:29 05/05/04
Go up one level in this thread
On May 05, 2004 at 10:13:27, Rolf Tueschen wrote: >"Jan Selschotter - Heidi Wyffels" <j.NOselschotterSPAM> wrote: > >> "Robert Hyatt" <hyatt@crafty.cis.uab.edu> schreef in bericht >> news:aks1th$ffv$2@juniper.cis.uab.edu... >>> marco <vialospamm.ledro@tiscali.it> wrote: >>> > Hi at all >>> > I want change my PC.... >>> > What is the Crafty SMP performance improvement compare to same processor >>> > family/tipe ? ( for example in Kn/sec or in elo points ) >>> > Is there a site with some data table ? >>> > THX >>> > Marco >>> >>> I am not sure what you are asking. If you mean "how much faster will >>> crafty run on a dual processor at XXXXmhz per cpu than it will run on >>> a single XXXXmhz cpu?" then the answer is about 1.7 times faster on >>> average. Sometimes more, sometimes less... >> And on N cpu's? Does speed(N)/speed(N-1) go to 1 for big N? > >I suspect so but there has been no way to test it so far. In a better >SMP algorithm used in Cray Blitz, for 16 cpus the speedup had flattened >to 11.1 from an optimal of 16.0. using 2 cpus it was very close to 2.0, >close enough that rounding to 1 decimel place made it 2.0. So obviously >the curve is hyperbolic and has an asymptote somewhere out there. And I >doubt it gets to 32 for any number of processors... > >Crafty is not as efficient, so it might have a max of 16 or worse, since >I have not tested it on a large SMP box (yet). > >For one thru four processors, crafty's SMP speedup is about > >speedup = 1 + .7 * (NCPUS -1) > >which gives about 3.1 for 4 cpus. Whether it will hold for 8 and >16 I don't know (yet). > > > > > > >-- >Robert Hyatt Computer and Information Sciences >hyatt@cis.uab.edu University of Alabama at Birmingham >(205) 934-2213 115A Campbell Hall, UAB Station >(205) 934-5473 FAX Birmingham, AL 35294-1170 Thanks but you should not post this kind of stuff. Using actual facts in a discussion with Vincent is not appropriate. Fiction is the name of the game. The fact that I have posted the above dozens of times doesn't matter. BTW I did change it later after testing on 8 processors, as 8 seems to fit the general formula I gave OK. But no real data beyond 8 so far, except for a few runs by Eugene when we were studying the Itanium NUMA results (and I should point out that the above specifically mentions a SMP platform, not NUMA so that would not change SMP results anyway...
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.